Head to head: MoonshotAI: Kimi K3 vs Anthropic: Claude Fable 5.1

MoonshotAI: Kimi K3: # KiteBridge Software Bug Triggers 13,400 Duplicate Reimbursements Worth $2.8 Million *Expense-management vendor halts automated payments, recovers most funds, and faces questions about a safeguard it claimed to have built two versions ago.* A failed safeguard inside KiteBridge's expense-management software sent $2.8 million in duplicate reimbursements to the employees of 91 corporate customers on Tuesday, forcing the company to suspend its automated payment system and spend the next day and a half clawing the money back. By Wednesday evening, KiteBridge said it had recovered $2.3 million, with the remaining $500,000 still pending. The failure began with a routine deployment. At 8:45 a.m. Tuesday, KiteBridge pushed version 4.2 of its platform live — the kind of incremental update software-as-a-service companies release constantly. The first customer complaint arrived at 9:12 a.m., less than half an hour later. By 10:03 a.m., roughly 80 minutes after the release went out, the company had suspended automated payments entirely. In that short window, the software generated 13,400 duplicate reimbursements. ## A retry that looked like a new request According to the company, the fault lay in how the updated software handled a payment-processor timeout. KiteBridge does not move money itself; it hands reimbursement instructions to an outside processor that executes the transfers. When that processor timed out — failing to confirm whether a payment had gone through — KiteBridge's fallback system began retrying, a standard response. The flaw was in what happened next: instead of recognizing each retry as a second attempt at the same transaction, the system treated it as an entirely new reimbursement request and paid it again. The effect was not unlike a cashier who, unsure whether a card charge succeeded, keeps swiping — with every swipe going through. The episode carries a particular sting because KiteBridge claimed to have solved exactly this problem. In an April changelog, the company advertised that version 4.0 had introduced "exactly-once payment protection" — an industry term for guaranteeing that a transaction executes a single time, even if attempted repeatedly. Tuesday's duplicates suggest that protection either failed or was undermined by the 4.2 release. ## Slow to say "duplicate" The company's public communication lagged its private knowledge. KiteBridge's incident page initially described the problem as delayed reimbursements — money arriving late, when in fact it was arriving multiple times. The page did not disclose the duplicate payments until six hours after automated payments were suspended. The company did stress that no employees had money withdrawn from their own accounts: the duplicate payments were funded by employers, whose corporate accounts front reimbursement money. ## Recovery and unanswered questions In a statement, CEO Lena Ortiz apologized to customers and said KiteBridge would publish a technical report on the incident. That report will have to answer uncomfortable questions: why a safeguard marketed two versions ago failed on Tuesday, why the system did not flag thousands of anomalous payments before customers did, and why the public incident page described the opposite of the actual problem for six hours. Recovering the remaining $500,000 may prove the harder task. Because the money landed in employees' bank accounts as apparently legitimate reimbursements, KiteBridge and its customers will likely have to request that overpayments be returned — a process that depends partly on recipients' cooperation. For a company whose product amounts to a promise that corporate money moves correctly, the cost of Tuesday's bug will be measured in more than dollars.

Anthropic: Claude Fable 5.1: # Software Flaw at KiteBridge Sends $2.8 Million in Duplicate Reimbursements to Employees A routine software update at expense-management company KiteBridge went wrong within half an hour of release on Tuesday, generating 13,400 duplicate reimbursement payments worth $2.8 million across 91 corporate customers. By Wednesday evening the company had clawed back $2.3 million, with the remaining half-million dollars still pending. The duplicates were funded by employers, not employees. No worker had money pulled from a personal account, but companies that rely on KiteBridge to pay back staff for travel, meals and other business expenses found their reimbursement budgets drained twice over for the same claims. ## A fallback that did the opposite of its job KiteBridge deployed version 4.2 of its platform at 8:45 a.m. Tuesday. The first customer complaint landed 27 minutes later, at 9:12 a.m. The failure began with something mundane: a timeout from KiteBridge's payment processor, the outside service that actually moves money to bank accounts. Timeouts are common in payment systems, and software is typically built to retry a transaction when it does not get a timely response. KiteBridge's system did retry. The problem was in how its fallback logic interpreted those retries. Instead of recognizing them as second attempts at the same payment, it treated each one as a brand-new reimbursement request. Every stalled transaction spawned a fresh payment, and the original often went through as well. That is precisely the scenario the company had claimed to have engineered away. An April changelog announcing version 4.0 said the release introduced "exactly-once payment protection," a safeguard meant to guarantee that a given reimbursement is paid a single time no matter how many retries occur behind the scenes. Tuesday's incident suggests that protection either did not cover the fallback path or was undermined by changes in version 4.2. KiteBridge suspended all automated payments at 10:03 a.m., roughly 80 minutes after the deployment and 51 minutes after the first complaint. By then the duplicates had already been issued. ## A slow and incomplete public account The company's handling of its public communications has drawn additional scrutiny. KiteBridge's incident status page initially described the problem as *delayed* reimbursements, language that implied employees were waiting on money rather than receiving it twice. It was not until six hours after automated payments had been halted that the page disclosed the duplicate payments. For much of Tuesday, affected finance teams had no official confirmation that their accounts had overpaid, even as some were already fielding questions from employees who noticed two identical deposits. ## Recovery and promises of transparency Recovering the money has proceeded quickly in relative terms. Because the funds went to employees of KiteBridge's corporate customers rather than to unknown third parties, most could be reversed or offset through the employers' own payroll and expense channels. The $2.3 million recovered by Wednesday evening represents about 82 percent of the total; the balance remains pending. CEO Lena Ortiz apologized for the incident and said KiteBridge would publish a technical report explaining what happened. The company has not yet said when that report will be released, whether version 4.2 has been rolled back, or when automated payments will resume. For a company whose core product is the accurate movement of other people's money, the episode raises an uncomfortable question. KiteBridge had told customers in April that duplicate payments were a solved problem. Eight months later, a single processor timeout proved otherwise, and the company's first public explanation understated what had gone wrong. Customers will be watching the promised report closely to see how both failures are addressed.

MoonshotAI: Kimi K3: # Arcwell's Relay drafts replies for support agents — and flags the ones who might cancel Arcwell has launched Relay, an artificial intelligence platform that reads customer support conversations and suggests answers to the agents handling them, adding the company to a growing field of AI tools aimed at customer service teams. The pitch: fewer hours lost to searching, and earlier warnings when a customer is about to walk. Relay is built to work alongside agents, not replace them. It follows conversations as they unfold and drafts proposed replies. It identifies customers who appear at risk of canceling and alerts their agents. And it can compress long support threads into summaries, sparing agents the chore of writing up each interaction. The launch arrives with early results from a pilot program. Twelve companies tested Relay over six months, and Arcwell says agents using the platform closed tickets 14% faster. The company did not release customer-retention results — a conspicuous gap for a product whose pitch includes catching customers before they leave. Relay costs $65 per agent per month, so the bill grows with the size of the support team. Chief executive Priya Shah framed the product as a fix for a familiar inefficiency. "Support agents spend too much time searching for information that their companies already have," she said. Relay's answer is to bring that information to the agent — its suggestions are drawn from the same internal material agents would otherwise hunt through themselves. That access is also the catch. To do its job, Relay requires a company's support transcripts and internal help documents — sensitive material, and a level of access some buyers may hesitate to grant. Arcwell says customer data is encrypted and is not used to train shared models, and companies can choose to retain their data for 30 days or 90. For now, the hardest evidence behind Relay is the 14% speed gain from Arcwell's own pilot. The number that may matter more to buyers — how many at-risk customers the platform actually keeps — is one the company has yet to publish.

Anthropic: Claude Fable 5.1: # Arcwell's Relay Puts an AI Assistant Beside the Support Desk Arcwell has introduced Relay, a platform that reads customer support conversations as they happen and feeds agents suggested replies, in a bid to cut the time staff spend hunting for answers their companies already hold. The software performs three jobs. It drafts responses for agents to send or adapt. It flags customers whose conversations suggest they may be about to cancel. And it writes summaries of exchanges, sparing agents the chore of typing up notes after each ticket. "Support agents spend too much time searching for information that their companies already have," said Priya Shah, Arcwell's chief executive. The pitch rests on a six-month trial with 12 companies. According to Arcwell, agents using Relay closed tickets 14% faster than they otherwise would have. That is the only performance figure the company has published. It has not released results for the cancellation-prediction feature, which leaves open the question of whether Relay actually helps businesses keep customers or merely tells them who is about to leave. Relay is priced at $65 per agent per month. To do its work, the platform needs broad access to a company's internal material — specifically, its support transcripts and help documentation. That requirement will give some prospective buyers pause, since support logs routinely contain account details, complaints and other sensitive information. Arcwell says the data is encrypted and is not used to train models shared across customers, meaning one client's conversations should not shape the suggestions another client sees. Companies can also set how long Relay holds their data, choosing between a 30-day and a 90-day retention window. Those assurances address the most obvious objections, but they are Arcwell's own, and the company has not described any independent audit of its handling of customer data. The launch places Arcwell among a growing number of vendors applying generative AI to customer service, where the volume of repetitive questions and the abundance of written records make the field an obvious target. Whether Relay distinguishes itself will depend on evidence the company has yet to provide: how the 14% figure holds up outside a controlled pilot, and whether its churn warnings translate into customers who stay.

MoonshotAI: Kimi K3: # Raincheck raises $8 million for flood-risk software *The Austin startup helps cities decide which drainage problems to address before storms.* Raincheck has raised an $8 million seed round led by Basin Ventures. ## Expansion plans The company plans to hire eight engineers.

Anthropic: Claude Fable 5.1: # Raincheck raises $8 million for flood-risk software *The Austin startup helps cities decide which drainage problems to address before storms.* Raincheck has raised an $8 million seed round led by Basin Ventures. ## Expansion plans The company plans to hire eight engineers.

MoonshotAI: Kimi K3: # Cloudnote Left Prompt Excerpts From 1,842 User Sessions Exposed for Four Hours *The company removed the debugging page 19 minutes after a researcher flagged it and says it has found no evidence anyone else saw the data. But its investigation is not finished — and the lapse raises questions about how customer prompts are handled.* Cloudnote, a software company with 3.2 million registered users, confirmed that a debugging page containing excerpts of customer prompts was accessible without authentication for four hours on Monday. Company logs show the page held text from 1,842 user sessions, with each excerpt capturing as many as 200 characters of a prompt — enough to contain a sentence or two of whatever a user had typed. Cloudnote said full documents, passwords and payment information were stored elsewhere and were not present on the page. A security researcher discovered the page and notified the company, which removed it 19 minutes later. Cloudnote said it has found no evidence that anyone other than the researcher accessed the page during the four-hour window. That assurance comes with a caveat: the company's investigation is not complete. Among the questions still unanswered are how the page became publicly accessible in the first place and whether Cloudnote will notify the users whose prompts appeared in its logs. Even a 200-character fragment can carry sensitive material. Users routinely entrust writing and productivity apps with personal messages, medical details and confidential business information, and the episode raises a pointed question: why were live customer prompts flowing through a debugging page at all?

Anthropic: Claude Fable 5.1: # Cloudnote Says Debugging Page Exposed Prompt Excerpts From 1,842 User Sessions *The page was reachable without a login for about four hours on Monday. The company says it has found no sign anyone but the researcher who reported it looked, but its investigation is not finished.* Cloudnote said a debugging page containing excerpts of customer prompts was accessible without authentication for roughly four hours on Monday. The company removed the page 19 minutes after a security researcher notified it of the exposure. Server logs show the page displayed excerpts from 1,842 user sessions, according to Cloudnote. Each excerpt could include up to 200 characters of a prompt, enough in some cases to indicate what a user was working on or asking the service to do. Full documents, passwords and payment information were stored elsewhere and were not present on the page, the company said. Cloudnote said it has found no evidence that anyone other than the researcher accessed the page while it was exposed. That finding is preliminary: the company has not completed its investigation, and it has not said what caused the page to be left open or when it expects to finish its review. The affected sessions represent a small fraction of Cloudnote's 3.2 million registered users. The incident still carries weight because even short fragments of prompts can reveal personal, professional or confidential details, and because a page meant for internal debugging was exposed to the public internet at all. What remains unknown is whether the logs capture every visit and how many distinct users the 1,842 sessions represent.

MoonshotAI: Kimi K3: # After seven years and 212 startups, Foundry Hall will close its doors *The Milwaukee accelerator's largest sponsor declined to renew. Six employees will lose their jobs, and twelve companies must find new space.* Foundry Hall, the nonprofit startup accelerator that has anchored Milwaukee's early-stage business community since 2018, will close on October 31, executive director Camille Ross said. The organization has enough money to operate through October but not enough to run another full program. The immediate cause was the loss of its largest corporate sponsor. Blank River Bank, which supplied 38 percent of Foundry Hall's annual budget, declined to renew its sponsorship. A bank spokesperson said the decision reflects a shift in the bank's community investment priorities and was unrelated to the accelerator's performance. Inside the building, that distinction landed as little more than a technicality. Ross said her team spent much of the past year approaching foundations, manufacturers, and other banks, hoping to replace the funding before the current program ended. None of those conversations produced a commitment in time. Winding down now, she said, was the only option that let the organization finish the current program on its feet. "We kept waiting for the replacement check that never came," she said. The numbers the accelerator leaves behind are not small. Over seven years, Foundry Hall supported 212 startups, which now employ a collective 1,140 people. Its pitch to founders was simple: workspace, mentorship, and introductions to investors under one roof. Its closure will eliminate six staff positions and displace the twelve startups that currently work out of the building. For founders like Dev Malik, the loss is less about square footage than about what the building made possible. "My first investor meeting happened in that building," Malik said. "Before Foundry Hall, I was pitching from the back table of a coffee shop." City officials say they are trying to soften the landing. Deputy Mayor Susan Cole said the city has begun speaking with building owners and local universities about space for the twelve displaced companies, though no arrangement has been finalized. Whether any option will be ready before the end of October remains unclear. Ross said the staff will spend its final weeks helping resident companies relocate rather than recruiting a new class. The accelerator's last day will fall roughly seven years after it opened. Malik said he plans to keep working from the space until the end. After that, he isn't sure where his next investor meeting will be.

Anthropic: Claude Fable 5.1: # Foundry Hall to Close After Losing Its Largest Sponsor The lights at Foundry Hall will go out on October 31. The nonprofit startup accelerator, which opened in Milwaukee in 2018, announced this week that it will shut down at the end of the month after its largest corporate sponsor, Blank River Bank, declined to renew its support. The bank had supplied 38 percent of the organization's annual budget. Foundry Hall has enough cash to keep its doors open through October. It does not have enough to run another full program, and its leadership chose to close rather than start a cohort it could not finish. Over six years, the accelerator supported 212 startups. Those companies now employ 1,140 people across the region. Executive director Camille Ross said the organization spent months trying to fill the hole left by the bank's exit, approaching other sponsors, foundations and individual donors. None came through at the scale required. "We kept waiting for the replacement check that never came," Ross said. In a statement, Blank River Bank described the decision as a change in its community investment priorities and said it was unrelated to Foundry Hall's performance. The bank did not say where that money will go instead. The closure will cost six employees their jobs. It will also displace the twelve startups currently working out of the accelerator's shared space. Dev Malik is not among them. His company came through the program years ago, and he has since moved on to his own offices. But the announcement landed hard anyway. "My first investor meeting happened in that building," Malik said. "Before Foundry Hall, I was pitching from the back table of a coffee shop." He said the hardest part was not nostalgia for his own time there. It was knowing that the founders inside now were mid-stride, and that the floor was about to disappear beneath them. City officials say they are trying to soften the landing. Deputy mayor Susan Cole said the city is speaking with building owners and local universities about space for the twelve companies before the end of the month. She did not offer a timeline, and nothing has been signed. Ross is clear-eyed about what that help can and cannot do. Desks can be found. Wi-Fi can be installed. What is harder to replace is the thing Foundry Hall actually offered, which was never square footage. It was a room where someone with an idea and no track record could sit across from someone who had already been through it, and be taken seriously. The twelve current companies have until October 31 to pack up. Ross said staff will spend the final weeks helping them find new addresses, connecting them with mentors who have agreed to keep meeting informally, and closing out the books. After that, the building goes back to its landlord, and 212 companies will have an alma mater that no longer exists.

MoonshotAI: Kimi K3: **C** Option C is the strongest lede because it accurately conveys the key facts: what happened (product pulled), when (Tuesday), what was exposed (workspace titles, not message contents), how many accounts (4,860), and how long (11 hours). The others fail because: - **A** falsely claims messages were leaked and uses loaded language ("privacy disaster"). - **B** reads like PR spin ("innovative," "difficult decision") and omits the actual news—the data exposure. - **D** is factually wrong: message contents were not sent, and there's no evidence of hacker access.

Anthropic: Claude Fable 5.1: C

MoonshotAI: Kimi K3: # Two AI Models Tied for Best Customer Support. One Costs 61% Less. A new benchmark of five language models on real customer-support work has a clear takeaway for teams shopping for automation: the two best models are statistically indistinguishable on quality — but one of them is dramatically cheaper and meaningfully faster. Researchers ran 200 anonymized customer-support tickets through five models — North, Vale, Ember, Kite and Harbor — and had each produce one proposed answer per ticket. Three experienced support managers graded every response blindly, meaning they didn't know which model wrote which answer. Scores combined factual correctness, usefulness and tone into a single average. ## A tie at the top, a chasm on price Model North posted the highest average score at 84.2, with Model Vale close behind at 83.8. That half-point gap sounds like a ranking, but it isn't: the researchers found the difference was not statistically significant. In practical terms, the two models tied. What wasn't close was the cost. North's estimated cost came in at $1.90 per 1,000 tickets. Vale's was $0.74 — roughly 61% cheaper. Vale also answered 18% faster. For a company handling tens of thousands of tickets a month, choosing the statistically tied-but-cheaper model is one of the easier procurement decisions imaginable. ## The rest of the field falls into clear tiers Below the leaders, the benchmark separates into distinct quality bands rather than a smooth gradient. Model Ember averaged 79.1 at $0.31 per 1,000 tickets — a meaningful step down in quality for a meaningful step down in price. Model Kite (76.4, $0.22) and Model Harbor (71.0, $0.08) continued the slide, with Harbor's rock-bottom price buying the weakest answers by a wide margin. The pattern here matters more than any individual number: cost tracks quality across the field, but with one big exception at the top. You generally get what you pay for — except that Vale gives you North-tier quality at a mid-tier price. ## Different strengths suggest a routing strategy The averages hide an interesting split. North performed best on billing disputes, while Vale performed best on technical troubleshooting. That suggests the choice between the two leaders may not be either/or. A support operation could route ticket types to the model that handles them best — billing to North, technical issues to Vale — though doing so would mean paying North's premium only for the slice of volume where it genuinely leads. ## One weakness all five models share The clearest red flag in the results applies to every model tested: all five performed poorly when a ticket depended on an image attachment. Screenshots of error messages, photos of damaged products — anything visual degraded performance across the board. Any deployment of these models for frontline support will need a human fallback or a separate workflow for image-dependent tickets, which are common in real support queues. ## Read these results with care Several limitations bound how far these findings travel. The 200 tickets were English-language and came from a single software company, so the rankings may not hold for other industries, tones of voice or customer bases. The test measured only single, written answers — researchers did not test agentic actions, tool use or multilingual support, all of which increasingly define what AI support means in production. A model that drafts a good reply may still fail at issuing a refund, querying an account database or switching languages. And 200 tickets, while enough to separate clear tiers, leaves smaller gaps — like the half point between North and Vale — within the noise. ## The bottom line If your tickets look like these, the benchmark makes a strong case: Vale delivers top-tier quality at roughly 40% of the leader's cost, with faster responses to boot. But treat the category strengths and the shared weakness on images as seriously as the headline averages. The best model on paper is only the best model for your queue if its strengths match the tickets you actually get — and if you've planned for the ones none of them can handle.

Anthropic: Claude Fable 5.1: # The Cheaper Model Matched the Expensive One. That's the Story. When a research team put five language models to work on 200 real customer-support tickets, the headline wasn't that one model won. It was that the top two finished in a statistical dead heat — and one of them cost less than half as much to run. Model North posted the highest average score, 84.2. Model Vale came in at 83.8. That 0.4-point gap sounds like a ranking, but the researchers checked and found it wasn't statistically significant. With 200 tickets and three human graders, that difference is indistinguishable from noise. Grade the same tickets again next week and Vale could just as easily land on top. What *is* distinguishable is the bill. North's estimated cost was $1.90 per 1,000 tickets. Vale's was $0.74. Vale also returned answers 18% faster. For a support operation handling tens of thousands of tickets a month, that's the same quality for roughly 40 cents on the dollar, delivered sooner. ## How the test worked The setup was deliberately simple. Researchers collected 200 anonymized tickets from a single software company — all in English — and fed each one to five models. Every model produced one proposed answer per ticket, so there was no cherry-picking across multiple attempts. Three experienced support managers then graded the answers without knowing which model had written what. Their scores blended three things: Was the answer factually correct? Was it actually useful to the customer? Did it strike the right tone? Combining these into a single number is a judgment call, but it reflects how a support lead would evaluate a human agent's draft, which is the point. Blind grading by practitioners is the methodological strength here. These weren't automated metrics or model-graded outputs; they were people who do this job deciding whether the answer would have helped. ## Where the models differed Below the top pair, the field spread out more clearly. Ember, Kite, and Harbor scored 79.1, 76.4, and 71.0 respectively, at progressively lower prices. Harbor, at $0.08 per 1,000 tickets, is almost free — but a 13-point gap from the leaders is large enough to mean something. Whether it means "unacceptable" depends entirely on what the answers are for. A model drafting suggestions for a human agent to edit has a lot more room for error than one replying to customers directly. The more interesting detail is that the two top models didn't excel at the same things. North performed best on billing disputes; Vale performed best on technical troubleshooting. If a company's queue skews heavily toward one category, the "statistically tied" verdict could tilt. A SaaS firm drowning in configuration questions might find Vale pulls ahead; a subscription business fielding refund arguments might lean toward North despite the cost. ## The shared weakness One finding cut across every model: when a ticket depended on an image attachment — a screenshot of an error, a photo of a damaged item — all five performed poorly. The models weren't failing because one was worse at images than another; they were failing because the task fundamentally required looking at something they couldn't adequately interpret, or because the test pipeline didn't give them the picture at all. Either way, any team deploying these tools should expect to route image-dependent tickets to humans, or at least flag them for review. ## What this doesn't tell you The limitations are real and worth taking seriously: - **One company, one language.** All tickets came from a single software vendor and were written in English. A model's performance on a bank's multilingual queue could look very different. - **Answers only.** The models were asked to write replies. They were not asked to look up account details, issue refunds, or take any action. Agentic workflows and tool use were out of scope, and those are increasingly how support automation is being built. - **A single pass.** Each model produced one answer per ticket. No retries, no prompt tuning, no system-prompt experiments. Real deployments optimize heavily, and models respond to optimization differently. - **200 tickets.** Enough to separate a 71 from an 84. Not enough to separate an 83.8 from an 84.2 — which the researchers correctly acknowledged. ## The practical takeaway The benchmark doesn't crown a best model. It does something more useful: it shows that paying the most isn't buying measurably better answers, at least on this kind of work. For most teams, the question isn't "North or Vale?" but "how much quality do I need, and what's the cheapest model that reliably delivers it?" On this evidence, Vale is the obvious place to start that conversation — with a human still in the loop for anything with a screenshot attached.

MoonshotAI: Kimi K3: # HelioDock Launches Night Shift, Remote Monitoring for Warehouse Robot Fleets *Tested for 14 months with nine logistics companies and 310 robots, the software will cost $90 per robot per month — and it arrives alongside a $24 million Series B led by North Coast Ventures.* HelioDock, a maker of autonomous warehouse robots, on Thursday launched Night Shift, software that lets operators monitor robot fleets remotely — and announced $

Anthropic: Claude Fable 5.1: # HelioDock Launches Remote Robot Monitoring Software, Raises $24 Million *Night Shift lets warehouse operators oversee autonomous fleets from off site, but the system hands control back to the robots when the network goes down.* HelioDock, a maker of autonomous warehouse robots, on Thursday launched Night Shift, software that lets operators monitor and manage their robot fleets remotely. The company also announced a $24 million Series B round led by North Coast Ventures, bringing its total funding to $39 million. The release targets a familiar problem in automated warehouses: a single stalled robot that pulls a manager on site after hours. HelioDock says Night Shift reduced unplanned robot stoppages among its beta customers by 17%, a figure that has not been independently verified. One early user said the software performed well where connectivity was strong but lost much of its value during network outages. "A warehouse manager should not have to drive across town at two in the morning because one robot is confused by a pallet," CEO Mara Velez said in announcing the product. ## Fourteen months in private beta Night Shift spent 14 months in a private beta that involved nine logistics companies and 310 robots. According to HelioDock, customers in that program saw 17% fewer unplanned stoppages compared with operating without the software. The company did not say how the figure was measured, and no outside party has confirmed it. Aaron Pike of Fulcrum Logistics, one of the beta customers, said the system worked well in facilities with strong connectivity. In locations that relied on rural LTE, he said, it became less useful whenever the cellular connection dropped. That limitation follows from how the product is built. Night Shift adds a remote layer on top of the robots' existing programming rather than replacing it. When connectivity fails, the robots return to their locally configured behavior, and remote control becomes unavailable until the connection is restored. A robot that is stuck on a pallet when the network drops will behave as it would have before Night Shift was installed. For operators with reliable wired or cellular networks, that fallback may rarely come into play. For facilities in areas with spotty coverage, Pike's experience suggests the software's benefit will track the quality of the connection. ## New capital and a crowded field The $24 million Series B led by North Coast Ventures more than doubles the capital HelioDock has raised to date. The company did not disclose how it plans to allocate the money. HelioDock is not alone in selling fleet-management tools for warehouse robots. Relay Robotics and YardOS both offer competing products. HelioDock did not compare Night Shift to those offerings in its announcement, and the company's figures from the beta program were not measured against rival systems. ## Pricing and availability Night Shift will cost $90 per robot per month. For a fleet the size of the entire beta program, 310 robots, that works out to roughly $27,900 a month. The software becomes generally available September 15. HelioDock has positioned the product around the cost of after-hours site visits. Whether a $90 monthly fee per robot pays for itself will depend on how often a given facility's robots stall, how far managers currently travel to reset them, and how reliably the facility stays connected. Velez's framing of the problem, a manager driving across town at two in the morning, describes the scenario the software is meant to eliminate. Pike's account from the beta describes the scenario it cannot: a robot that needs help at a moment when no one can reach it. Between those two cases sits the question that operators evaluating Night Shift will have to answer for their own buildings.

MoonshotAI: Kimi K3: # Raincheck Raises $8 Million to Help Cities Get Ahead of Flood Risk *The Austin startup's software pairs drainage maps with rainfall forecasts so local governments can deploy maintenance crews before the rain arrives.* Raincheck, an Austin startup that makes flood-risk software for local governments, has raised $8 million in seed funding, the company announced today. The company maps drainage systems and combines those maps with rainfall forecasts, an approach detailed in [Raincheck's published methodology](https://example.com/raincheck-methodology). Cities use the software to decide where to send maintenance crews before a storm. Basin Ventures led the round, with participation from existing investor Long Mile. Raincheck has now raised $10.5 million in total. CEO Ana Flores said the software is built to solve a problem of prioritization: > "Our customers usually know which streets flood. The harder problem is deciding which blocked drain, broken pump or unfinished repair should receive attention before the rain arrives." Raincheck currently works with seven cities in Texas and Louisiana, though the company declined to identify them because some contracts have not been publicly announced. Raincheck's forecasts are recommendations; emergency managers retain control over deployments and public warnings. The company's competitors in flood planning include HydroMap and municipal engineering contractors. Raincheck said the new funding will help it hire eight engineers and expand into Florida.

Anthropic: Claude Fable 5.1: # Raincheck Raises $8 Million to Help Cities Clear Drains Before the Storm *The Austin startup's software pairs drainage maps with rainfall forecasts to tell local governments where to send maintenance crews first.* Raincheck, an Austin company that makes flood-risk software for local governments, announced today that it has raised $8 million in seed funding. The round was led by Basin Ventures, with participation from existing investor Long Mile. The company has now raised $10.5 million in total. Raincheck maps municipal drainage systems and combines those maps with rainfall forecasts. Cities use the software to decide where to send maintenance crews before a storm arrives. The company has published [an overview of how its flood-risk model works](https://example.com/raincheck-methodology). CEO Ana Flores said the challenge for cities is not knowing where water will collect, but choosing among competing fixes with limited time: > Our customers usually know which streets flood. The harder problem is deciding which blocked drain, broken pump or unfinished repair should receive attention before the rain arrives. Raincheck currently works with seven cities in Texas and Louisiana. It declined to identify them, saying some contracts have not been publicly announced. The company's forecasts are recommendations only. Emergency managers retain control over crew deployments and public warnings. Raincheck competes with other flood-planning vendors, including HydroMap, as well as municipal engineering contractors. The company says the new funding will go toward hiring eight engineers and expanding into Florida.

MoonshotAI: Kimi K3: # QueueZero Launches Proxy, an AI Assistant That Turns Meeting Talk Into Assigned Work *At Wednesday's launch event, an offhand remark from the audience became two tasks assigned to CEO Evan Wu.* **SAN FRANCISCO** — QueueZero on Wednesday launched Proxy, an AI meeting assistant that can attend meetings, produce notes and assign follow-up work to software agents. The event included an unscripted demonstration of what that last feature means in practice. As QueueZero CEO Evan Wu presented onstage, a customer in the audience said, "Someone should tell Evan to stop talking and send us the pricing sheet." Proxy, which was listening to the proceedings, interpreted the remark as an instruction. It created two tasks, both assigned to Wu: "Stop talking" and "Send pricing sheet." Wu laughed when the tasks appeared, completed the second one and deleted the first. QueueZero said Proxy behaved as designed. The assistant requires user confirmation before it sends any external message, the company said. Internal tasks, by contrast, can be created automatically, on the reasoning that they are low-stakes and easily removed. Administrators can also disable automatic task creation entirely. The company is positioning that design as a safeguard. A misread instruction that generates an internal to-do item can be deleted in a click, as Wu demonstrated. A misread instruction that emails a client is a different category of problem, and QueueZero said the confirmation step exists to prevent it. Proxy enters a crowded market. AI transcription and summarization have become standard features across meeting and collaboration software in recent years, offered by large platform vendors and startups alike. QueueZero's bet is that the next step is delegation: converting conversation into assigned work, handed off not only to people but to software agents that can send documents, update records and complete other routine tasks. Wednesday's incident illustrated both the promise and the unresolved difficulty of that approach. Proxy correctly extracted actionable items from casual speech — speech that was not addressed to it and was never intended as a command. Distinguishing an instruction from an aside remains one of the harder problems for AI systems built to act on what they hear, and it is a problem that gets harder, not easier, in real meetings, where people routinely say things they do not mean. Wu completed the "Send pricing sheet" task himself. The sheet went out; the talking continued. Whether the company's controls are sufficient will be tested at scale in the coming weeks. Proxy costs $30 per user per month and enters public beta next week.

Anthropic: Claude Fable 5.1: # QueueZero's New Meeting Assistant Turns an Audience Aside Into a Task List QueueZero launched Proxy, an AI meeting assistant, on Wednesday, and the product's first public demonstration offered an unplanned illustration of both its capabilities and the questions surrounding them. Proxy is designed to join a meeting on a user's behalf, generate notes and assign follow-up work to software agents. The company positions it as a step beyond transcription tools: rather than simply recording what was said, Proxy is meant to identify commitments and turn them into actionable items without human intervention. That design was tested midway through the launch event. While QueueZero CEO Evan Wu was walking through the product's feature set, a customer in the audience said, "Someone should tell Evan to stop talking and send us the pricing sheet." Proxy, which was running live during the demonstration, interpreted the remark as an instruction. It created two tasks and assigned both to Wu: "Stop talking" and "Send pricing sheet." Wu laughed, completed the second task and deleted the first. He then continued the presentation, which ran to its scheduled length. The exchange drew a brief laugh from the room, but it also highlighted a central tension in the product category. Tools that act autonomously on spoken language must distinguish between a directive and a passing comment, and the two are not always separated by anything more than tone. QueueZero says Proxy includes safeguards for the higher-stakes case. The system requires explicit confirmation from a user before any message is sent outside the organization, meaning an offhand remark cannot result in an email to a client or vendor without someone signing off. Internal tasks, however, can be created automatically, which is how the two items appeared on Wu's list during the demonstration. Administrators have the option to disable automatic task creation entirely, the company said, in which case Proxy will surface proposed tasks for review rather than adding them directly. The company has not said whether it plans to adjust how Proxy interprets indirect or third-person statements. The audience member's comment was addressed to the room generally rather than to Wu or to the software, a distinction the system did not make. Proxy is priced at $30 per user per month. It enters public beta next week, giving a wider group of customers the opportunity to see how it handles the particular grammar of their own meetings. The incident was minor and quickly resolved. But it offered a concrete example of what customers are being asked to trust: a system that listens to everything said in a room and decides, on its own, which parts were meant to be done.

MoonshotAI: Kimi K3: # Startup Spotlight: LedgerLark Takes On Post-Trade's Longest Days *The startup's AI agents hunt down trade mismatches for small investment firms — and leave the final call to humans.* The most expensive problems in finance are not always the glamorous ones. Ask Mina Batra. During seven years in operations at a regional brokerage, she watched highly paid staff burn days on work that boiled down to finding where two numbers stopped agreeing. One incident stuck: a six-person team spent three days resolving a single trade mismatch that turned out to be caused by a copied account number. Most exceptions, Batra observed, were not exotic. They were typos, timing gaps and duplicated entries — errors that hide easily in high volumes and consume hours wildly disproportionate to their complexity. That observation became the starting point for LedgerLark, which she co-founded with Owen Carr, who previously built reconciliation systems at a large bank. The pairing is practical: Batra knows the workflow from the inside, and Carr knows the plumbing. ## What the agents actually do LedgerLark's software reads three sources of record — trade confirmations, custodian records and the firm's internal ledgers — and compares them. When the records diverge, its AI agents investigate the discrepancy and propose a correction, which a human staff member must approve before anything changes. The scope is notably narrow, and that seems intentional. LedgerLark does not execute trades and does not move money. That boundary should lower the risk and compliance hurdles for adoption, and it gives cautious operations teams a reason to say yes: the agents do the legwork, people keep the authority. For a small investment firm, where the operations team might be a handful of people, each unresolved exception is a direct claim on someone's evening. ## Early traction, with caveats Five firms are paying customers, on annual contracts ranging from $36,000 to $110,000. The spread suggests pricing that flexes with firm size or scope — and a company still working out where its value lands. LedgerLark says one customer reduced its unresolved daily exceptions from 70 to 18. If that result holds across clients, it would represent a meaningful cut in the loose ends operations teams carry overnight. But it is a single, company-reported data point, measured over an unspecified period, from a customer base of five. The company has raised a $3.5 million seed round led by Motive Row — enough to fund the next phase, though not an answer to the harder questions. ## The crowded middle LedgerLark has company in this market: established reconciliation vendors and the internal software large banks build for themselves. Incumbents bring distribution, existing integrations and years of institutional trust, and there is little stopping them from adding agent-style features to products firms already own. LedgerLark's counter is focus — small investment firms are often too small to justify enterprise reconciliation suites and too stretched to build their own. Adoption may be slow regardless. Financial institutions typically require lengthy security and procurement reviews, which can stretch sales cycles for a young vendor — and stretch the patience of a five-customer revenue base. The company will also need to show that its agents' suggestions are right often enough that reviewing them is genuinely faster than doing the work from scratch, and to learn how customers' auditors treat agent-proposed corrections. ## What to watch Near-term signals: whether customers six through ten arrive faster than the first five; whether contract values drift toward the top of the range; how incumbents respond. The underlying observation — that post-trade reconciliation at small firms is understaffed, repetitive and expensive — is hard to dispute. Whether AI agents are the right fix, and whether LedgerLark is the team to deliver it, is now a matter of evidence.

Anthropic: Claude Fable 5.1: # Startup Spotlight: LedgerLark Wants to Catch the Errors That Slip Through Post-Trade Most of the attention in financial technology goes to the moment a trade happens. LedgerLark is interested in what comes after: the confirmations, custodian statements and internal ledgers that have to agree with one another before anyone can say with confidence that a trade settled correctly. The company builds AI agents for post-trade operations at small investment firms, the kind that often lack the dedicated technology budgets of larger institutions but face many of the same reconciliation chores. ## Where the idea came from Co-founder Mina Batra spent seven years in operations at a regional brokerage. The moment she points to as the origin of LedgerLark was not dramatic so much as grinding. A six-person team spent three days resolving a single trade mismatch. The cause, once found, was a copied account number. That experience, as she describes it, raised a simple question: why did a problem that small take that long to locate? The answer had less to do with carelessness than with how the work is structured. Operations staff compare records across systems that were never designed to talk to each other, and the comparison itself is manual. Her co-founder, Owen Carr, approached the same problem from the opposite side. He built reconciliation systems at a large bank, where the tooling exists but is often tailored to one institution's workflows and difficult to adapt elsewhere. ## What the product does LedgerLark's agents read three sources, trade confirmations, custodian records and a firm's internal ledger, and look for places where they disagree. When the agents find a discrepancy, they propose a correction and route it to a person for approval. The boundary is deliberate. The product does not execute trades and does not move money. Humans retain the final decision, which keeps the software on the advisory side of a line that regulators and compliance teams care about. It also limits the damage if an agent reasons incorrectly, since a bad suggestion is only a suggestion. Whether that boundary holds as customers ask for more automation is an open question. Approving dozens of proposed corrections a day can itself become routine, and routine approval is not the same as review. ## Early traction Five firms are paying customers, with annual contracts ranging from $36,000 to $110,000. The company says one customer reduced its unresolved daily exceptions from 70 to 18. That figure comes from LedgerLark rather than an independent audit, and it describes one firm, so it is best read as an encouraging data point rather than a benchmark. The company has raised a $3.5 million seed round led by Motive Row. ## The harder parts LedgerLark is not entering an empty field. Established reconciliation vendors have served this market for years, and larger institutions have built internal software that does much of the same matching. Small firms may be underserved by those options, but underserved is not the same as unserved, and incumbents can move toward smaller clients if the segment proves worthwhile. Sales cycles are another constraint. Financial institutions commonly require lengthy security and procurement reviews before adopting software that touches trade data. For a company with five customers and a seed-stage budget, each of those reviews consumes time and attention. Five paying firms is a real signal, but it is also a small sample from which to generalize. There is also the question of how much of the problem AI agents actually solve. Reconciliation errors often originate upstream, in data entry or in mismatched reference data. Catching them faster is valuable. Preventing them is a different product. ## What to watch The near-term test for LedgerLark is whether its first customers renew and whether the exception reductions hold up across firms with different custodians and record formats. If the agents generalize, the company has a plausible path. If each new client requires substantial customization, it may find itself rebuilding the same bespoke systems its founders left behind.

MoonshotAI: Kimi K3: 6

Anthropic: Claude Fable 5.1: **6**